ChatGPT’s Code Interpreter can produce sophisticated data analyses with a click, how to use AI as a brainstorming partner, and more, in this week’s roundup of AI-related links—curated especially for healthcare leaders.

• ChatGPT’s newest feature is a seriously big deal. Wharton professor Ethan Mollick writes about his early experiments with ChatGPT’s Code Interpreter, which can create and run computer code and statistical analyses on demand. His prompts include “prove to a doubter that the earth is round with code” and “here is some data on superhero powers, look through it and tell me what you find.” Mollick’s takeaway: Code Interpreter “is the single most useful, interesting mode of AI I have used.”
• What AI can—and can’t—do for human doctors. Writing in the New York Times, physician Daniela J. Lamas shares her evolving view of the role of AI in medicine. She suggests AI will find patterns and make diagnoses that humans might miss, but also argues it “will never replace a hand at the bedside, eye contact, understanding—what it is to be a doctor.”
•  GPT-4 can (sometimes) diagnose illnesses that baffle doctors. A new JAMA research letter shows that GPT-4 can offer state-of-the-art differential diagnoses for tricky medical case studies.
• AI hasn’t displaced human translators. Yet. Despite predictions that advances in machine translation could end the need for human translators, the industry is still humming along—albeit with big changes, such as lower per-word rates and higher pre-translator productivity. The trends may foretell what’s coming for healthcare and other industries as AIs develop new skills.
• How to use AI as a brainstorming partner. From Harvard Business Review, a roundup of ways to use AI to augment your creativity. Among the practices: promoting divergent thinking, challenging expertise bias, assisting in idea evaluation, supporting idea refinement, and facilitating collaboration.
• Don’t make these mistakes in your AI prompts. You’ve heard of “prompt engineering”: the skill of writing effective LLM prompts. But how do these best practices compare to the ways real-world users interact with AI? This paper identifies several common user errors, such as failing to give examples of their desired output. One striking tidbit: “A number of participants made assumptions about the kinds of instructions LLMs respond well to after just one attempt … overgeneralizing from a single (typically negative) example.”
• Hire an AI to torment your telemarketers. On the lighter side, here’s a Wall Street Journal piece on an AI chatbot that wastes telemarketers’ time with halfway-relevant inanities. When one marketer asks how much a man owes on his credit cards, for instance, the AI replies, “I have so many of them, you know. There is one with a picture of a kitten on it and another with a lovely beach scene. Do you like kittens or beaches?”